1 / 37

Jeremy C. W. Chan*, Qian Ding*, Patrick P. C. Lee , Helen H. W. Chan

Parity Logging with Reserved Space: Towards Efficient Updates and Recovery in Erasure-coded Clustered Storage. Jeremy C. W. Chan*, Qian Ding*, Patrick P. C. Lee , Helen H. W. Chan The Chinese University of Hong Kong FAST’14. The first two authors contributed equally to this work.

espen
Télécharger la présentation

Jeremy C. W. Chan*, Qian Ding*, Patrick P. C. Lee , Helen H. W. Chan

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript

  1. Parity Logging with Reserved Space: Towards Efficient Updates and Recovery in Erasure-coded Clustered Storage Jeremy C. W. Chan*, Qian Ding*, Patrick P. C. Lee, Helen H. W. Chan The Chinese University of Hong Kong FAST’14 The first two authors contributed equally to this work.

  2. Motivation • Clustered storage systems provide scalable storage by striping data across multiple nodes • e.g., GFS, HDFS, Azure, Ceph, Panasas, Lustre, etc. • Maintain data availability with redundancy • Replication • Erasure coding

  3. Motivation • With explosive data growth, enterprises move to erasure-coded storage to save footprints and cost • e.g., 3-way replication has 200% overhead; erasure coding can reduce overhead to 33% [Huang, ATC’12] • Erasure coding recap: • Encodes data chunks to create parity chunks • Any subset of data/parity chunks can recover original data chunks • Erasure coding introduces two challenges: (1) updates and (2) recovery/degraded reads

  4. Challenges 1. Updates are expensive • When a data chunk is updated, its encoded parity chunks need to be updated • Two update approaches: • In-place updates: overwrites existing chunks • Log-based updates: appends changes

  5. Challenges 2. Recovery/degraded reads are expensive • Failures are common • Data may be permanently loss due to crashes • 90% of failures are transient (e.g., reboots, power loss, network connectivity loss, stragglers) [Ford, OSDI’10] • Recovery/degraded read approach: • Reads enough data and parity chunks • Reconstructs lost/unavailable chunks

  6. Challenges • How to achieve both efficient updates and fast recovery in clustered storage systems? • Target scenario: • Server workloads with frequent updates • Commodity configurations with frequent failures • Disk-based storage • Potential bottlenecks in clustered storage systems • Network I/O • Disk I/O

  7. Our Contributions • Propose parity-logging with reserved space • Uses hybrid in-place and log-based updates • Puts deltas in a reserved space next to parity chunks to mitigate disk seeks • Predicts and reclaims reserved space in workload-aware manner  Achieves both efficient updates and fast recovery • Build a clustered storage system prototype CodFS • Incorporates different erasure coding and update schemes • Released as open-source software • Conduct extensive trace-driven testbed experiments

  8. Background: Trace Analysis • MSR Cambridge traces • Block-level I/O traces captured by Microsoft Research Cambridge in 2007 • 36 volumes (179 disks) on 13 servers • Workloads including home directories and project directories • Harvard NFS traces (DEAS03) • NFS requests/responses of a NetApp file server in 2003 • Mixed workloads including email, research and development

  9. MSR Trace Analysis Distribution of update size in 10 volumes of MSR Cambridge traces • Updates are small • All updates are smaller than 512KB • 8 in 10 volumes show more than 60% of tiny updates (<4KB)

  10. MSR Trace Analysis • Updates are intensive • 9 in 10 volumes show more than 90% update writes over all writes • Update coverage varies • Measured by the fraction of WSS that is updated at least once throughout the trace period • Large variation among different workloads • need a dynamic algorithm for handling updates • Similar observations for Harvard traces

  11. Objective #1: Efficient handling of small, intensive updates in an erasure-coded clustered storage

  12. Saving Network Traffic in Parity Updates Linear combination with some encoding coefficient • Make use of linearity of erasure coding • CodFS reduces network traffic by only sending parity delta Question: How to save data update (A’) and parity delta (ΔA) on disk? A B C P applying the same encoding coefficient Update A P A’ B C P’ is equivalent to P’ A’ parity delta (ΔA)

  13. Update Approach #1: in-place updates (overwrite) • Used in host-based file systems (e.g., NTFS and ext4) • Also used for parity updates in RAID systems Update A Disk Disk A B C B C Problem: significant I/O to read and update parities

  14. Update Approach #2: log-based updates (logging) • Used by most clustered storage systems (e.g. GFS, Azure) • Original concept from log-structured file system (LFS) • Convert random writes to sequential writes Update A Disk Disk A B C A B C Problem: fragmentation to chunk A

  15. Objective #2: Preserves sequentiality in large read (e.g. recovery) for both data and parity chunks

  16. Parity Update Schemes Our Proposal O: Overwrite L: Logging

  17. Parity Update Schemes Data stream FO FL PL PLR Storage Node 1 Storage Node 2 Storage Node 3

  18. Parity Update Schemes Data stream a b FO a b a+b FL a b a+b PL a b a+b PLR a b a+b Storage Node 1 Storage Node 2 Storage Node 3

  19. Parity Update Schemes Data stream a a’ b FO a b a+b FL Δa a a’ b a+b PL Δa a b a+b PLR a b a+b Δa Storage Node 1 Storage Node 2 Storage Node 3

  20. Parity Update Schemes Data stream a a’ c b d FO c c+d d a b a+b FL Δa a a’ b a+b c d c+d PL Δa a c b a+b c+d d PLR a b a+b c Δa c+d d Storage Node 1 Storage Node 2 Storage Node 3

  21. Parity Update Schemes Data stream a a’ c b’ b d FO c c+d d a b a+b FL Δa Δb a a’ b b’ a+b c d c+d PL a c b a+b Δb Δa c+d d PLR a b a+b c Δa c+d d Δb Storage Node 1 Storage Node 2 Storage Node 3

  22. Parity Update Schemes Data stream a a’ c b’ c’ b d FO c c+d d a b a+b FL Δa Δb Δc a a’ c’ b b’ a+b c d c+d PL Δc a c b a+b Δb Δa c+d d PLR a b a+b c Δa c+d d Δb Δc Storage Node 1 Storage Node 2 Storage Node 3

  23. Parity Update Schemes Data stream FO: extra read for merging parity a a’ c b’ c’ b d FL: disk seek for chunk b FO c c+d d a b a+b FL Δa Δb Δc a a’ c’ b b’ a+b c d c+d FL&PL: disk seek for parity chunk b PL Δc Δa a c b a+b Δb c+d d PLR a b a+b c Δa c+d d Δb Δc PLR: No seeks for both data and parity Storage Node 1 Storage Node 2 Storage Node 3

  24. Implementation - CodFS • CodFS Architecture • Exploits parallelization across nodes and within each node • Provides a file system interface based on FUSE OSD: Modular Design

  25. Experiments • Testbed: 22 nodes with commodity hardware • 12-node storage cluster • 10 client nodes sending • Connected via a Gigabit switch • Experiments • Baseline tests • Show CodFS can achieve theoretical throughput • Synthetic workload evaluation • Real-world workload evaluation Focus of this talk

  26. Synthetic Workload Evaluation Random Write Logging parity (FL, PL, PLR) helps random writes by saving disk seeks and parity read overhead FO has 20% less IOPS than others IOZone record length: 128KB RDP coding (6,4)

  27. Synthetic Workload Evaluation Sequential Read Recovery merge overhead No seeks in recovery for FO and PLR Only FL needs disk seeks in reading data chunk

  28. Fixing Storage Overhead PLR (6,4) • FO (8,6) is still 20% slower than PLR (6,4) in random writes • PLR and FO are still much faster than FL and PL FO/FL/PL (8,6) FO/FL/PL (8,4) Recovery Random Write Data Chunk Parity Chunk Reserved Space

  29. Dynamic Resizing of Reserved Space • Remaining problem • What is the appropriate reserved space size? • Too small – frequent merges • Too large – waste of space • Can we shrink the reserved space if it is not used? • Baseline approach • Fixed reserved space size • Workload-aware management approach • Predict: exponential moving average to guess reserved space size • Shrink: release unused space back to system • Merge: merge all parity deltas back to parity chunk

  30. Dynamic Resizing of Reserved Space Step 1: Compute utility using past workload pattern disk shrink previous usage current usage smoothing factor disk Step 2: Compute utility using past workload pattern write new data chunks disk shrinking reserved space as a multiple of chunk size avoids creating unusable “holes” no. of chunk to shrink Step 3: Perform shrink

  31. Dynamic Resizing of Reserved Space Reserved space overhead under different shrink strategies in Harvard trace 16MB baseline Shrink only performs shrinking at 00:00 and 12:00 each day Shrink + merge performs a merge after the daily shrinking *(10,8) Cauchy RS Coding with 16MB segments

  32. Penalty of Over-shrinking Average number of merges per 1000 writes under different shrink strategies in the Harvard trace Penalty of inaccurate prediction Less than 1% of writes are stalled by a merge operation *(10,8) Cauchy RS Coding with 16MB segments

  33. Open Issues • Latency analysis • Metadata management • Consistency / locking • Applicability to different workloads

  34. Conclusions • Key idea: Parity logging with reserved space • Keep parity updates next to parity chunks to reduce disk seeks • Workload aware scheme to predict and adjust the reserved space size • Build CodFS prototype that achieves efficient updates and fastrecovery • Source code:http://ansrlab.cse.cuhk.edu.hk/software/codfs

  35. Backup

  36. MSR Cambridge Traces Replay PLR ~ PL ~ FL >> FO Update Throughput

  37. MSR Cambridge Traces Replay PLR ~ FO >> FL ~ PL Recovery Throughput

More Related